A BAYESIAN METHOD FOR OSCILLATOR CHARACTERIZATION

Transcription

1 30th Annual Precise Time and Time Interval (PTTI) Meeting A BAYESIAN METHOD FOR OSCILLATOR CHARACTERIZATION F. Vernotte Observatoire de Besanson 41 bis av. de l'observatoire, BP 1615, F25010 Besanson, France tel $33: fax francoisqobsbesani,on.fr G. Zalamansky Universitd de Metz Campus Bridowr, F57000 Metz, France Abstract The power. spectral density of frequ,ency fluctuation,^ of an oscillator is generally modeled us u surrt of power laws with integer exponent.? (from 2 to +2). However, a power law with a fractional elcponenl: may exist. We propose a method for measurin,g such a noise level and determining the probability density of the exponent. This yields a criterion, for compatibility with, an integer exponent. This method is based upon a Bayesian approach called the reference analysis of BernardoBerger. The application to a sequence of frequency measurement from a quartz oscillator illustrates this puper. INTRODUCTION It is cornrnonly assumed that S,(f). the power spectral density (PSD) of frequency deviation of an oscillator, may be modeled as the surrl of 5 power laws, defining 5 types of noise: +? s',(f) = C h,fa (1) a=2 where h, is the level of the f" rloise. But it may be noticed that models with noninteger exponents are occasionally used. The estimation of the noise levels is mainly achieved by using the Allan variance [ll. which is defined versus the integration tirne T as: = 2 ((&+I q kl2) (a 99

2 Report Documentation Page Form Approved OMB No Public reporting burden for the collection of information is estimated to average 1 hour per response, including the time for reviewing instructions, searching existing data sources, gathering and maintaining the data needed, and completing and reviewing the collection of information. Send comments regarding this burden estimate or any other aspect of this collection of information, including suggestions for reducing this burden, to Washington Headquarters Services, Directorate for Information Operations and Reports, 1215 Jefferson Davis Highway, Suite 1204, Arlington VA Respondents should be aware that notwithstanding any other provision of law, no person shall be subject to a penalty for failing to comply with a collection of information if it does not display a currently valid OMB control number. 1. REPORT DATE DEC REPORT TYPE 3. DATES COVERED to TITLE AND SUBTITLE A Bayesian Method for Oscillator Characterization 5a. CONTRACT NUMBER 5b. GRANT NUMBER 5c. PROGRAM ELEMENT NUMBER 6. AUTHOR(S) 5d. PROJECT NUMBER 5e. TASK NUMBER 5f. WORK UNIT NUMBER 7. PERFORMING ORGANIZATION NAME(S) AND ADDRESS(ES) Observatoire de Besanson,41 bis av. de l observatoire,bp 1615, F25010 Besanson, France, 8. PERFORMING ORGANIZATION REPORT NUMBER 9. SPONSORING/MONITORING AGENCY NAME(S) AND ADDRESS(ES) 10. SPONSOR/MONITOR S ACRONYM(S) 12. DISTRIBUTION/AVAILABILITY STATEMENT Approved for public release; distribution unlimited 11. SPONSOR/MONITOR S REPORT NUMBER(S) 13. SUPPLEMENTARY NOTES See also ADA th Annual Precise Time and Time Interval (PTTI) Systems and Applications Meeting, Reston, VA, 13 Dec ABSTRACT see report 15. SUBJECT TERMS 16. SECURITY CLASSIFICATION OF: 17. LIMITATION OF ABSTRACT a. REPORT unclassified b. ABSTRACT unclassified c. THIS PAGE unclassified Same as Report (SAR) 18. NUMBER OF PAGES 9 19a. NAME OF RESPONSIBLE PERSON Standard Form 298 (Rev. 898) Prescribed by ANSI Std Z3918

3 I fli I variance versus the integration time r is plotted, the graph exhibits different slopes, each slope corresponding to a type of noise: (r) = &(,(s) = h ~f" and ~=~l, (3) The estirrlation of C',, yields an estirrlation of the noise level h,. However, this curve may exhibit an exponent p which seems to be noninteger. Does this mean that the corresponding PSD is not compatible with the 5 power law model? In this paper, we propose a method for estimating the rnost probable value of this exponent in order to solve this ambiguity. This method is applied to an example of stability measurement. 1 CLASSICAL STABILITY ANALYSIS OF AN OSCILLATOR Sequence of frequency measures 1 Figure 1 shows average frequency measures i& of a 10 MHz quartz oscillator compared to a c.esium clock. The sampling rate is 10 s and the integration time of each frequency I measure is also 10 s (sampling without dead time). In order to obtain dimensionless y, samples, we must subtract the nominal frequency (10 MHz) from the frequency measures and normalize by vo: Variance analysis Figure 2 is a loglog plot of the Allan deviation of the quartz y, samples versus the integration time r. A least squares fit of these variance measures (solid line), weighted by their uncertainties, detects only two types of noise: a white noise and an fp%oise. The corresponding noise level estimations are: ho = (2.2 f 0.4). 1 K5s at IF (68% confidence) h2 = (2.3 f 0.6). l~''s~ at la (68% confidence) (for the assessment of the h, noise levels and their uncertainties, we used the multivariance method described in PI). However, for large T values (corresponding to low frequencies), the variance measures move away from the fitted curve. Two explanations are possible:. instead of an fp2 noise, there is a noise whose noninteger exponent is contained between 2 and 3 ; since the uncertainty domains of the variance measures contain the fitted curve, this apparent divergence may be due to a statistical effect. In order to choose between these two explanations, we decided to estimate the probability density of the exponent with a Bayesian approach.

4 BAYESIAN APPROACH The goal of all measurement is the estimation of an unknown quantity 6' from measures $, i.e. determining p(81$), the density of probability of the quantity B knowing the measures $. The Bayesian theory is based on the following equality [31: ~(40 pt<l0)74@) where p(ei6') is the distrihutiorl of the measures < for a fixed value of the quantity 6' and r(6') is the a priori density of probability of the quantity 0, i.c. before performing any measurement. The determination of this a priori density, called the prior, is generally one of the main difficulties of this approach (particularly in the case of total lack of knowledge! ). In this paper, we use the Jeffrey's prior which ensures properties slich as invariance [31. Spectral density and covariance matrix Let us define the vector y whose components are the N jj,, samples. We assume that y is s Gaussian vector. The probability distribution of y is: where C: is the covariance matrix. Since &(f) is the Fourier transform of R,(T), the autocorrelation function of the frequency deviation. the general term of C: is: Equation (7) reveals the key role played by the spectral density of the noise in the expected fluctuation. We will present a general rr~ethod for estimating the parameters of the model for S,(f). Assumed model for the spectral density We assume that the sequence of frequency measures is composed of a white noise y, whose variance (i.e. the level) is unitary, and of a red noise y, whose level is unknown, multiplied by the real standard deviation of the white noise u,,, (0, is easily estimated from high sampling rate frequency ~aeasurernent): y = (h, + yr)rw. This yields the following model for Sy(f): s,(f) = ho +he.f" with 3<a<1 where h,o = 2.2. lojs, h, and a are the unknown parameters. Let us denote y, the normalized vector: Yn "Yw +YT. 101

5 The corresponding rlorrrlalized PSD S,( f) is: where U, is an amplitude factor whosc meaning will be explained below (see equation (23)). Statistical model The part of the spectral density due to the red noise y, may bc written: We used the BernardoBerger analysis e = (0, H). Construction of the estimat0r.s: 41 for estimating the unknown parameter Let us introduce the orthonormal basis of!xn, {po,...,pj,...,p~~}, defined such as the ith component of pj is: Pi,? = Fj (fi) (13) where ti is the date of the ith frequency measure and &(t) is a polynomial of degree j: satisfying the orthonormality condition 151: C Fj (ti).f.(ti) = djk. It can be shown than the scalar product of a vector pj by the noise vector y is an estirrlate of the noise spectrurn for a given frequency fj [GI. Let us denote Cj such an estimate applied to the normalized noise: Practically, we limited to 16 the number of estimators pj (from degrees 0 to 15) for lim iting the computation and because the high degrees, estimating the high frequencies, are less informative for a red noise. Moreover, in order to ensure convergence for very low frequencies (even if the low cutoff frequency tends towards O), the polynomials must satisfy the moment condition [k 61: the minimum degree jmi, of a polynomial to ensure convergence up to an exponent cr is: a 1 jmin 2 2 ' (16) Since we have assumed a 5 3, the first 2 estimators and must be removed. Thus we have n = 14 estimators {pz,..., yls) and n = 14 estimates {&,...,ti5). Construction of the priors: The covariance matrix defined in relationship (7) is an ensemble average of the different,estimate products over an infinite number of realizations of this process:

6 As for the noise vector y, the estimate vector $ may be split into two terms, according to equations (10) and (15): The covariarlce matrix may also bc split: C'= t = 411, + 5r (19) (<w<:) + ( ETfl) = In th.lln. V(@) where I, is the identity matrix in!rn and the general term of the matrix V(a) is: The high cutoff frequency fh, duration of the sequence. in (21) is the Nyquist frequency and T is the total Let E*(CY) denote the ith eigenvector of V(a) and ;/,(a) its ith eigcnvalue (i E (0,..., n 1)). The averaged quadratic norm of the estimate vector ( is: The factor u, is chosen in such a way that. for H = 1. thc avcraged quadratic norms (~l ~ll~) and (llk2) are eqllal: The direct problem is now solved since < is a vector of Vbith a, probability distribution given the parameter B equal to: Denoting "Tr(M)" the trace of a matrix and X the matrix defined as: the Fisher information matrix I(@) is (see M): The Jeffrey's prior ~(6) is defined as: The parameter B is a twodimensional parameter composed of the exponent parameter cu and of the amplitude parameter H. Since we are rrlostly interested in a, H is called a nuisance parameter. In presence of nuisance parameter, Bernardo and Berger suggested that a should first be fixed and the conditional prior ~(111cw) cornpilted for that value. The full prior is then: T(@) = ~ (Hja). ~ ( L Y ). (28)

7 The conditional prior x(hla) is given by: 4Hla) = m (29) whcrc [1(0)]11, [I(#)llz = [I(#)Jzl and [I(B)]22 are the elements of the Fisher information matrix I(O). The prior for a may be computed as: where c is a normalization coefficient ensuring that J n(a)da = 1 and: This prior is plotted in Figurc 3. Construction of the posteriors: According to the Bayes theorem, the posterior probability distribution is given by: The posterior probability distribution for cr is then given by: RESULTS AND DISCUSSION Compatibility with an integer exponent Figtire 4 shows the posterior probability distribution for the exponent a of the red noise using the BernardoBerger prior. The exponent value obtained for the maximum of likelihood, just as for the maximum of the distribution, is a = 2.2. However, a = 2 is fully compatible with this prior distribution. Thus, we may conclude that the apparent divergence between the variance measures and the fitted curve in Figure 2 is probably due to a statistical bias of the data. The spectral density S,(f) is then compatible with the following model: Noise level estimation Selecting an exponent value a = 2, we obtained the posterior probability distribution plotted in Figure 5. As in the variance analysis, we chose a confidence interval of 68% (16% probability that hvz is smaller than the low bound and 16% probability that hl2 is greater than the high bound): hl2 = ( 2.3 ' dlg ). 1 0l~~' at 1u (68% confidence)

8 The difference between the maximum likelihood value (122 = s1) and the variance analysis value (hpz = lo'%') is only 0.18%. However, the confidence intervals given by these two methods are quiet different. The main difference concerns the symmetry of thc variance analysis interval: i11 this case, we don't take into account the fact that the noise levels are positive, whereas the prior of the Bayesian approach is null for negative values of hpz. Moreover, the variance analysis interval seems to be a bit underestimated. CONCLUSION The variance analysis is an useful tool for a quick estimation of the noise levels in the output signal of an oscillator. However, a negative estimate of a noise level nlay occur. Generally, in this case, this valuc is rejected and the corresponding noise level is assumed to be null. On the other hand. although the Bayesian method is a bit heavier, it takes into account properly the a priori information, and gives a more reliable estimation of these noise levels and especially of their confidence intervals. However, the main advantage of the Bayesian method concerns the verification of the validity of the power law model of spectral density. Each time the model is suspected, such an approach should be used in order to estimate the exponent of the power law. In particular, this method should be very interesting for the study of the fv1 and l ' f noise, whose origin remains mysterious [TI. REFERENCES [l] J. A. Barnes, A. R.. Chi, L. S. Cutler, D. J. Healey, D. B. Leesson, T. E. McCuningal,.J. A. Mullen, W. L. Smith, R. L. Sydnor, R. Vessot, and G. M. R. Winkler, 'cc~~,uraclerization of frequeracy stability, '' IEEE Transactions on Instrumentation and Measurement, IM20, [2] F. Vernotte, E. Lantz, J. Groslambert, and J. Gagnepain, "Oscillator noise analyszs: multiuariance measurement, '' IEEE Trans. on Inst. and Meas., IM42, [3] J. Bernardo and A. Smith, Bayesian theory, Wiley & Son, [4] G. Zalamansky and C. Robert, "Bayesian analysis with application to the timing residuals of a pulsar, " in Proceedings of MaxEnt'98, July 1998, Garching, Germany. [5] J. E. Deeter and P. E. Boynton, "Tcchniqucs for fh,e estimation of red power spectra. I. context and methodology," The Astrophysical Journal, 261, [6] F. Vernotte, G. Zalamansky, M. McHugh, and E. Lantz, "Cht08 frequencies and noise power law model of spectral densifg: adaptation oj the multivariance method for irregularly spaced timing data using thf lowest mode estimator approach.," IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control, 43, (3)) [7] M. Planat, V. Giordario, G. Marianneau? F. Vernotte, M. Mourey, C. Eckert, and J. Miehe, "Is the frequency noise of an, oscillator of deter.mini.stic origin P,'' IEEE 'I'rans. on U. F. F. C., 43, (2),

9 t I Time (s) Figure 1: Sequence of frequency measures 1 I 1 I I l l fitted variance response 1 f2... white noise...,",i '.., I I I I I 1 I I,, T (seconds) Figure 2: Allan deviation of t,he sequence of frequency measures

10 Figure 3: Reference prior for the power 0 Figure 4: Posterior probabilit,~ density for the power (I Figure 5: Posterior probability density for the noise level h,l2